acceptodds
Withdrawn from ICLR 2027

RoboBridge: A Self-Evolving Embodied Agent Framework for Sim-to-Real Transfer

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A key challenge in bringing embodied intelligence into the real world is transferring capabilities from simulation to reality and enabling agents to continually adapt after deployment. End-to-end vision-language-action policies provide strong manipulation capabilities, but their transfer to physical environments typically relies on calibrating simulated visual and dynamical conditions, collecting additional target-domain demonstrations, and optimizing the policy through further training. Tool-using embodied agents offer flexible task orchestration, yet existing systems primarily emphasize task execution and experience reuse within a given environment, with limited support for transferring procedural knowledge and continuously adapting it across simulation and reality. We propose RoboBridge, a framework that treats sim-to-real transfer as the continued adaptation of executable task skills. The agent represents task knowledge as procedures connecting task intent, observations, tool operations, and outcome verification. Interaction feedback is used to generate candidate skill revisions, which are evaluated before being persisted or rejected. A pretrained vision-language-action policy is exposed as a reusable action tool and enhanced with inference-time guidance, enabling fine-grained execution without retraining the underlying policy. RoboBridge grounds transferable skills in task semantics and interaction interfaces shared across simulation and reality. This representation preserves reusable task structure while allowing environment-dependent operations to be selectively revised through real-world execution feedback. We evaluate the framework on LIBERO-PRO and corresponding physical tasks, studying both skill evolution and post-transfer adaptation. Our framework provides a route from one-shot policy deployment to continual procedural learning across environments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Vision-Guided Manipulation, Large Language Model Agents, Vision-Language-Action Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xSnAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper descirbes an agentic VLM framework that maintains a library of skills. Each skill can either call a VLA or a position controller. The agentic system uses gathered experience from simulation or real world to mofidy and improve the skills that it executes. The framework is tested with GPT / codex and the pi05 VLA on the LIBERO-PRO benchmark. There are some further real-world validations over 3 pick-and-place tasks.

Critical strengths
  • The authors validated the framework in real-world robotic trials.
Critical weaknesses
  • This work presents a single data point: One specific VLM and one specific VLA where the proposed skill evolution leads to a small incremental improvement on one specific benchmark. There are no generalization tests over multiple VLMs, multiple VLAs, and there is no indication that these improvements span more than one benchmark. There are also no ablation studies which would provide useful insight on the design decisions of the skill evolution or guidance processes. The given ablation only shows that each of the two steps contributes to the overall improvement.
  • The above weakness unfortunately leads me to the conclusion that there is no scientific insight to draw from this submission in its current form. It is an interesting reference implementation, and the demonstrated incremental improvement motivates to look more deeply into this agentic system, but it may as well be that none of this works with a new version of GPT or some other VLA.
  • The title claims sim-to-real transfer, yet the method itself does not target this at all. It is simply that the abstraction of using two different checkpoints of pi05 allows for the high-level skills to transfer. It may be fair to claim "transferrable" skills, because the paper shows e.g. in Figure 5 that the evolved skills from simulation do transfer also to better success rate in real world, but the transfer itself is not achieved by the framework. Specifically, there is no indication that such transfer would work for a target real-world environment for which a suitable VLA checkpoint (trained on extensive environment specific human demonstrations) is not available. Further, it is not specified whether the 05 checkpoint used in simulation already has seen real-world training data during the VLA training (it would be surprising if it had not).
  • The skill guidance process is entirely unclear to me. It assumes a diffusion based action decoding that is incompatible with the definition that skills may freely choose from a VLA call and a position controller. This step seems only relevant to the VLA part, and only a specific subsets of VLAs uses diffusion for action decoding. This raises the question how modular the skills are after all. There is also no description of the reward function itself and its allows complexity. If the agent is free here to come up with any implementation, since it seems to be called in every denoising step, it can add significant computational overhead to the action decoding and may cause impractical inference times.
  • Section 3.2 appears to be critically unclear and gives me AI-writing-vibes. Examples of vague language:
    • does a skill specify "how a next operation should be selected" or does it specify which operation to select next? The way it is currently formulated a skill is nothing more than a prompt that then runs another LLM to select an action
    • skills are meant to generalize over "similar manipulation requirements" (line 237). It is not specified what these are, specifically as all the examples shown are pick and place. So is the entire framework one skill?
Questions for authors
  • What is the inference time for a single episode? How long does it take your real robot to put one cube into one bowl?
  • Is the action guidance limited to diffusion-based VLA decoders?
  • What is the claimed "structured scene description" (line 215)?
  • How is "observation freshness and uncertainty" measured (line 215)? Does the "uncertainty" have any probabilistic meaning? Is it processed in a probabilistic framework?
Reviewer_qXnzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RoboBridge combines an LLM planner, a fixed library of action tools, and pretrained frozen VLAs. Execution feedback is used to revise procedural skills, which are retained after validation. The agent can also configure reward-based guidance during action generation. Simulation-developed procedures are reused and refined on a real robot. Evaluation includes LIBERO-PRO, component ablations, and three real-robot tasks.

Critical strengths

The real-robot experiments show practical value in reusing simulation-developed procedures. The paper also includes disjoint training/validation/test seeds and component ablations, providing evidence that procedure revision, validation, and action guidance help within this system.

Critical weaknesses

Major

  1. The methodological advance over close prior work is unclear. Harness VLA [1] combines frozen VLAs, analytic tools, and memory-guided planning. ASPIRE [2, Sections 2.1–2.2 and 3.6] already uses execution traces to repair programs, validates reusable skills, and reuses simulation-discovered knowledge on real robots. Zetta [3, Sections 2.5–2.6] includes validation-gated updates, and VLS [4, Sections IV-A2 and IV-B2] steers frozen policies using VLM-generated differentiable rewards. These works are cited, but the paper does not clearly identify a new update rule, transfer mechanism, or guidance algorithm beyond their integration.

  2. The evidence for the integration's added value is limited. Success improves from Harness VLA's 72.1% to 74.3%, without reported variability across repeated runs or a sufficiently detailed account of matched observations, tools, checkpoints, and interaction/compute budgets. The ablations demonstrate useful components, but do not isolate an advantage over prior repair/guidance methods. Real-robot tests cover three tasks; the adaptation-cost comparison covers only one and needs to count simulation preparation costs. Since simulation and reality use different VLA checkpoints, these results primarily demonstrate procedural transfer rather than transfer of a shared motor policy.

Minor

The reward functions, action-to-reward computation, guidance settings, and actual before/after skill edits need more detail. Appendix E provides a schema and a schematic update record rather than a complete task-specific example.

[1] Zhang et al. Harness VLA. https://arxiv.org/abs/2607.08448 [2] Lu et al. ASPIRE: Agentic /Skills Discovery for Robotics. https://arxiv.org/abs/2607.00272 [3] Ding et al. Zetta. https://arxiv.org/abs/2608.16590 [4] Liu et al. VLS: Steering Pretrained Robot Policies via Vision-Language Models. https://arxiv.org/abs/2602.03973

Questions for authors
  1. Clarification: What concrete algorithmic differences distinguish your skill-update and transfer mechanisms from ASPIRE/Zetta, and your reward-guided action generation from VLS? Please provide an explicit update procedure and a complete reward example.
  2. Evidence affecting the assessment: With the same observations, tools, policy checkpoints, and interaction/compute budgets, how much improvement remains over Harness VLA and a comparable prior skill-repair system? What is the variability across repeated runs?
  3. Clarification/evidence: What advantage does the real-robot adaptation procedure provide over prior simulation-derived skill reuse, such as ASPIRE Section 3.6, when simulation preparation and real-world interaction costs are both counted?
Reviewer_iYmMOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

RoboBridge proposes an embodied-agent framework that combines a high-level coding agent, editable procedural skills, geometric action primitives, and frozen VLA policies. Execution feedback is used to propose additions, replacements, or deletions to a skill, with candidate revisions evaluated before being retained. The framework also introduces agent-configured reward guidance during action generation. The paper evaluates the framework on LIBERO-PRO and three real-robot tasks.

Critical strengths
  1. Transferring reusable procedural knowledge across environments is an interesting direction, particularly when retraining robot policies is costly. Separating task organization from low-level execution provides a useful perspective on this problem.

  2. Allowing existing knowledge to be revised or removed, while evaluating candidates before persistence, is a sensible approach to limiting the accumulation of unreliable experience. The ablations address several relevant components, including append-only updates, validation, and inference-time guidance.

Critical weaknesses
  1. The sim-to-real evidence is not convincing enough. The paper reports transferring skills, but does not show which simulation task produced each skill, what was transferred, or what was changed before real-world execution. Since the two environments use different VLA checkpoints, the same tool interface does not guarantee that a procedure will work with both policies. Comparing against no skill also leaves unclear whether the benefit comes from simulation experience or simply from having a reasonable task plan.

  2. The novelty over RPent/Harness VLA is unclear. Both use a coding agent to coordinate frozen VLAs and geometric tools, check execution results, and reuse procedural knowledge. The paper needs to explain more clearly what its skill revision and transfer mechanisms add beyond this existing framework.

  3. Key implementation details are missing. The paper does not explain how an agent-generated reward becomes a differentiable function for guiding actions. It also leaves the skill acceptance rule underspecified and provides no complete example showing an actual skill before and after revision.

  4. The experimental controls are limited. A static skill generated without simulation interaction would be an important baseline. Real-world results cover only three tasks with 15 trials per condition, and adaptation efficiency is tested on one task. Repeated runs and a clearer accounting of validation and evaluation costs are needed to establish the reliability of the reported gains.

Questions for authors
  1. Could the authors show the exact skill transferred for each real-world task, its source simulation task, and all changes made before the first physical trial?

  2. Does simulation-derived skill initialization outperform a task plan generated by the same LLM without simulation interaction, with all other components held fixed?

  3. What specifically distinguishes the method from RPent/Harness VLA, beyond describing procedural memory as skills?

  4. How are rewards implemented and differentiated, and what exact criterion determines whether a candidate skill replaces the current one?

  5. Do the reported adaptation costs include candidate validation and intermediate evaluation trials? How consistent are the results across independent runs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.